Reproduce “ImgCoT: Compressing Long Chain of Thought into Compact Visual Tokens for Efficient Reasoning of Large Language Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaoshu Chen, Sihang Zhou, Ke Liang, Taichun Zhou, Xinwang Liu
Researchers at the National University of Defense Technology introduce ImgCoT, a framework that compresses long textual Chain-of-Thought into compact visual tokens by rendering reasoning steps as structured images. This approach, and its hybrid variant Loose ImgCoT, enhance LLM reasoning efficiency and accuracy, consistently outperforming prior latent reasoning methods and often matching or exceeding full textual Chain-of-Thought while significantly reducing token consumption.
Key insights
- A "spatial inductive bias," derived from visual representations of CoT, is more effective for abstracting and preserving reasoning structure than the "linguistic inductive bias" of textual compression.
- Explicitly modeling logical dependencies within the visual rendering significantly contributes to the capture of structural relationships in compressed CoT.
- Combining abstract visual tokens with strategically chosen textual steps (L-ImgCoT) provides an optimal balance, addressing both global reasoning flow and fine-grained precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.